Unverified50% confidenceBenchmarkExact time
The skip-zero optimization (exploiting that 92.5% of neurons are zero after ReLU², reading down weights in groups of 16 neurons) raised speed from 6.6 tok/s to 7.3-7.6 tok/s, an improvement of about 12%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/17/2026
First Seen
Valid until: 1/15/2027
Sources
Related Entities
Related Claims
UnverifiedLeaky ReLU outputs a small proportion (usually 1%) of the input value for negative input, thereby avoiding the dead neuron problem.72% similarUnverifiedThe Φ value of standard feedforward neural networks approaches zero, because information flows only in one direction and lacks recurrent integration.67% similarUnverified30%–36% of the stable neurons for instruction-tuned generators are concentrated in BERT's layer 12, whereas for base models this proportion is below 14% in all cases.63% similarUnverifiedAfter introducing an explicit resolver to handle information supersession relationships, stale-memory failures dropped to zero and overall accuracy jumped from 33% to 78%.62% similarUnverifiedAccording to official performance tests, a search on a graph with 118,000 nodes saw response time drop from 24 milliseconds to 0.004 milliseconds, a speedup of approximately 6,000 times.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1053039API
curl https://kongchang.com/api/v1/knowledge/claims/1053039MCP
get_claim(id=1053039)